tars

Run a supervised autonomous loop on tars under explicit edit-scope and review-independence guardrails. Each cycle: preflight, propose, verify, halt-or-continue. Author and reviewer roles stay separate (producer-reviewer); the evaluator runs in a fresh sub-agent / fresh session and cannot self-certify. Contested cycles escalate to cross-vendor / multi-LLM review (Demerzel tribunal). Edits respect a line budget (rewrite-budget) so each diff stays small.

Self-improvement reflex. When the user corrects an approach ("no, don't do that", "we discussed this before", "stop X"), captures the lesson as a permanent project rule appended to CLAUDE.md so the pattern doesn't repeat in this or future sessions. Cherny called this "the most important loop" in his 2026 talk.

Capture meaningful session state (current cursor, in-flight work, live hypotheses, open questions, do-NOT-carry-forward, success criteria) to state/digests/latest.md so the next session — including one after auto-compaction — can re-enter without re-discovering context cold. Distinct from /learnings (which captures surprises). Validates against docs/contracts/digest-schema.json.